Colorectal cancer is the third most common malignancy worldwide and is highly curable when detected early. Traditional screening and diagnostic approaches include colonoscopy, stool-based tests, advanced imaging, and molecular assays. Machine learning and deep learning approaches are being applied across histopathology, endoscopy, imaging, and genomic data to improve detection, classification, and outcome prediction in colorectal disease.
This work introduces CDCTNet (colorectal diagnosis convolution transformer network), a hierarchical deep learning model that integrates convolutional neural network (CNN) blocks and a Vision Transformer (ViT) encoder with an interaction module to capture both local spatial detail and global contextual relationships for colorectal lesion diagnosis.
Recent literature demonstrates diverse DL applications in colorectal cancer: prediction of MSI/dMMR status from histology; mutation prediction such as KRAS from imaging; large-scale optical diagnosis models for colonoscopy; lymph node metastasis prediction from primary tumor slides; imaging-based response assessment; and biomarker extraction from whole-slide images. Reported performance across studies varies by task and dataset, and methods include ResNet variants, U-Nets, Faster R-CNN hybrids, and ensemble or weakly supervised frameworks. These works establish the potential and limitations of DL in CRC tasks, particularly the need to balance local morphological detail with broader contextual cues.
CDCTNet is a hybrid architecture designed to leverage complementary strengths of CNNs and transformers. The model consists of two parallel paths: a convolutional path comprising two convolution blocks to extract local high-dimensional spatial features, and a transformer path using a ViT encoder to model global correlations across the feature map. An IEM (information exchange module) facilitates interaction and fusion between these parallel feature representations. The fused features are pooled and classified with a softmax layer.
The convolutional branch contains two sequential convolution blocks. The first block (Conv1) applies 32 convolutional filters followed by a 2×2 max-pooling layer to produce local spatial feature maps. The second block (Conv2) applies 64 convolutions with another 2×2 max-pooling. ReLU activations are used after convolutions. The extracted spatial feature maps capture local texture and lesion morphology relevant to colorectal pathology.
A parallel ViT encoder is included to provide long-range feature correlation and global attention across the input representation. The ViT encoder processes embeddings derived from image patches to compute multi-head self-attention, enabling the model to capture contextual relationships that classical CNNs may miss. In CDCTNet, the ViT path operates in tandem with the convolutional path so that local and global information are both available for downstream classification.
The proposed IEM mediates interaction between convolutional spatial features and transformer embeddings. Spatial features extracted from the CNN branch are projected into the transformer embedding space via a dense projection. A learnable fusion parameter is used to optimize the relative contributions of CNN and ViT features before performing multi-head attention in the ViT encoder. This explicit information exchange is intended to strengthen attention to salient features and improve overall representation by combining local detail with global context.
The authors evaluated CDCTNet using the publicly available Kather and Kvasir datasets. The article reports standard classification metrics to assess performance, including precision, recall, F1 score, and Cohen’s Kappa. The dataset source is noted as a Zenodo repository in the manuscript. Details such as training hyperparameters, data splits, augmentation, optimizer settings, or cross-validation procedures were presented in the full manuscript but are not reproduced here in this summary beyond stating the datasets used and reported outcomes.
On the Kather and Kvasir datasets, CDCTNet achieved a reported precision of 96.60%, a Kappa score of 95.02%, recall of 98.08%, and an F1 score of 97.94%. These metrics indicate high agreement and strong sensitivity for the model on the evaluated test sets as reported by the authors.
The manuscript positions CDCTNet as a method that addresses two common limitations in prior approaches: the reliance of classical ML on handcrafted features and the limited global attention of standard CNNs. By combining CNN and ViT components with an IEM fusion strategy, CDCTNet aims to exploit local lesion morphology and broader contextual cues simultaneously. The paper situates this contribution within a broader literature of DL models for CRC, where tasks and datasets vary widely.
Limitations disclosed in the article include dataset scope (evaluation limited to the Kather and Kvasir datasets) and standard considerations for DL models such as data availability and model interpretability. The authors also note that data used in the study are publicly available and that they received no specific funding; no competing interests were declared.
CDCTNet is presented as a hierarchical hybrid network that combines convolutional feature extraction with a ViT encoder and an information exchange module to enhance feature attention for colorectal lesion diagnosis. Reported results on the Kather and Kvasir datasets show high precision, recall, F1, and Kappa values. The dataset used in the study is publicly accessible via Zenodo (link reported in the manuscript). The authors suggest that integrating local and global representations via explicit fusion can improve diagnostic performance; further validation on broader and prospective clinical datasets, as well as work on interpretability and deployment in clinical workflows, would be necessary next steps, though specific future work details beyond this summary were not reported in the source.